OpenAI says new safeguards blunt reasoning extraction on GPT-6.2
Original title: OpenAI makes it much harder for attackers to extract how its latest AI models think
Internal tests put prior-request extraction at 0.2 percent, down from 54.0 percent on GPT-6 Astra.
OpenAI said it tightened protections against extracting internal reasoning from GPT-6.2 Astra and GPT-6.2 Sol. In its automated tests, prior-request extraction fell from 54.0 percent on GPT-6 Astra to 0.2 percent, and current-request attacks from 22.9 to 0.9 percent. Defenses include training, input and output filters, API blocks, and weekly bans of thousands to tens of thousands of accounts. Selected improvements are also coming to GPT-6 Astra.
Unconfirmed: The success rates are OpenAI's own internal test figures, not independent measurements.
- Rates are aggregated across six attacker models in adaptive black-box tests.
- API blocks stop older models from reading protected reasoning.
- OpenAI bans thousands to tens of thousands of accounts weekly.
- Anthropic also reports banning accounts over distillation of internal reasoning.
- Vendors
- OpenAI
- Organizations
- Anthropic
- AI models
- GPT-6.2 AstraGPT-6.2 SolGPT-6 Astra
From the original article
1 min read at the-decoder.com
OpenAI has significantly strengthened safeguards against extracting internal reasoning from GPT-6. 2 Astra and GPT-6. 2 Sol.
By Matthias Bastian · the-decoder.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.